Huawei_Technology
/
MindSpore-AKG
Not watched
1
0
Fork 0
Code
Releases 1 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: efe42ee8fe
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'efe42ee8fe'
${ noResults }
MindSpore-AKG/tests/operators/gpu/test_all.py

545 lines
28 kB
Raw Blame History 

  1. #!/usr/bin/env python3

  2. # coding: utf-8

  3. # Copyright 2020-2021 Huawei Technologies Co., Ltd

  4. #

  5. # Licensed under the Apache License, Version 2.0 (the "License");

  6. # you may not use this file except in compliance with the License.

  7. # You may obtain a copy of the License at

  8. #

  9. # http://www.apache.org/licenses/LICENSE-2.0

  10. #

  11. # Unless required by applicable law or agreed to in writing, software

  12. # distributed under the License is distributed on an "AS IS" BASIS,

  13. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

  14. # See the License for the specific language governing permissions and

  15. # limitations under the License

  16. 

  17. from tests.common.test_utils import gen_random_shape

  18. 

  19. from tests.operators.gpu.test_ms_add import test_ms_add

  20. from tests.operators.gpu.test_ms_addn import test_ms_addn

  21. from tests.operators.gpu.test_ms_batch_matmul import test_ms_bmm

  22. from tests.operators.gpu.test_ms_exp import test_ms_exp

  23. from tests.operators.gpu.test_ms_maximum import test_ms_maximum

  24. from tests.operators.gpu.test_ms_minimum import test_ms_minimum

  25. from tests.operators.gpu.test_ms_mul import test_ms_mul

  26. from tests.operators.gpu.test_ms_divide import test_ms_divide

  27. from tests.operators.gpu.test_ms_rsqrt import test_ms_rsqrt

  28. from tests.operators.gpu.test_ms_sub import test_ms_sub

  29. from tests.operators.gpu.test_ms_tile import test_ms_tile

  30. from tests.operators.gpu.test_ms_one_hot import test_ms_one_hot

  31. from tests.operators.gpu.test_ms_sqrt import test_ms_sqrt

  32. from tests.operators.gpu.test_ms_cast import test_ms_cast

  33. from tests.operators.gpu.test_ms_reshape import test_ms_reshape

  34. from tests.operators.gpu.test_ms_expand_dims import test_expand_dims

  35. from tests.operators.gpu.test_ms_trans_data import test_ms_trans_data

  36. from tests.operators.gpu.test_ms_log import test_ms_log

  37. from tests.operators.gpu.test_ms_pow import test_ms_pow

  38. from tests.operators.gpu.test_ms_reduce_sum import test_ms_reduce_sum

  39. from tests.operators.gpu.test_ms_abs import test_ms_abs

  40. from tests.operators.gpu.test_ms_neg import test_ms_neg

  41. from tests.operators.gpu.test_ms_round import test_ms_round

  42. from tests.operators.gpu.test_ms_select import test_ms_select

  43. from tests.operators.gpu.test_ms_equal import test_ms_equal

  44. from tests.operators.gpu.test_ms_less_equal import test_ms_less_equal

  45. from tests.operators.gpu.test_ms_greater_equal import test_ms_greater_equal

  46. from tests.operators.gpu.test_ms_reciprocal import test_ms_reciprocal

  47. from tests.operators.gpu.test_ms_reduce_max import test_ms_reduce_max

  48. from tests.operators.gpu.test_ms_reduce_min import test_ms_reduce_min

  49. from tests.operators.gpu.test_fused_pad import test_fused_pad

  50. from tests.operators.gpu.test_fused_bn_reduce import test_fused_bn_reduce

  51. from tests.operators.gpu.test_fused_bn_update import test_fused_bn_update

  52. from tests.operators.gpu.test_fused_bn_follow_relu import test_fused_bn_follow_relu

  53. from tests.operators.gpu.test_fused_bn_follow_relu_avgpool import test_fused_bn_follow_relu_avgpool

  54. from tests.operators.gpu.test_fused_bn_double_follow_relu import test_fused_bn_double_follow_relu

  55. from tests.operators.gpu.test_fused_bn_reduce_grad import test_fused_bn_reduce_grad

  56. from tests.operators.gpu.test_fused_relu_grad_bn_reduce_grad import test_fused_relu_grad_bn_reduce_grad

  57. from tests.operators.gpu.test_fused_relu_grad_bn_double_reduce_grad import test_fused_relu_grad_bn_double_reduce_grad

  58. from tests.operators.gpu.test_fused_l2loss_grad import test_fused_l2loss_grad

  59. from tests.operators.gpu.test_fused_is_finite import test_fused_is_finite

  60. from tests.operators.gpu.test_fused_relu_grad_bn_update_grad import test_fused_relu_grad_bn_update_grad

  61. from tests.operators.gpu.test_fused_relu_grad_bn_double_update_grad import test_fused_relu_grad_bn_double_update_grad

  62. from tests.operators.gpu.test_fused_relu_grad import test_fused_relu_grad

  63. from tests.operators.gpu.test_fused_bn_update_grad import test_fused_bn_update_grad

  64. from tests.operators.gpu.test_fused_mul_div_rsqrt_mul_isfinite_red import test_fused_mul_div_rsqrt_mul_isfinite_red

  65. 

  66. 

  67. def add(poly_sch, fuzz_shape=None, mind_trick_str=''):

  68.     if fuzz_shape:

  69.         test_ms_add(fuzz_shape, fuzz_shape, 'float32', poly_sch=poly_sch)

  70.         return

  71. 

  72.     test_ms_add((1, 1024), (1, 1024), 'float32', poly_sch=poly_sch)

  73.     test_ms_add((2, 32, 256, 32, 32), (2, 32, 256, 32, 32),

  74.                 'float32', poly_sch=poly_sch)

  75. 

  76. 

  77. def addn(poly_sch, fuzz_shape=None, mind_trick_str=''):

  78.     test_ms_addn((1, 1024, 1024), "float32", 2, poly_sch=poly_sch)

  79.     test_ms_addn((1, 1024, 1024), "float16", 2, poly_sch=poly_sch)

  80. 

  81. 

  82. def bmm(poly_sch, fuzz_shape=None, mind_trick_str=''):

  83.     test_ms_bmm((768, 768), (768, 768), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  84.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  85.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 6", bind_thread="32 4")

  86.     test_ms_bmm((768, 768), (768, 768), 'float16', 'float16', layout1='NHTD', layout2='NHTD', layout_out='NHDT',

  87.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  88.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 6", bind_thread="32 8")

  89.     test_ms_bmm((32, 12, 128, 128), (32, 12, 128, 64), 'float16', 'float16', layout1='NHDT', layout2='NHTD', layout_out='NHDT',

  90.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  91.                 dim="0 0 1 1 0 1 1 1 0 2 64 64 0 3 64 64 0 4 64 4", bind_block="12 32", bind_thread="32 4")

  92.     test_ms_bmm((32, 12, 128, 64), (32, 12, 128, 64), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  93.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  94.                 dim="0 0 1 1 0 1 1 1 0 2 64 64 0 3 64 64 0 4 64 4", bind_block="12 32", bind_thread="32 4")

  95.     """ Bert Batch64

  96.     test_ms_bmm((8192, 768), (768, 768), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  97.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  98.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 64", bind_thread="32 4")

  99.     test_ms_bmm((8192, 768), (8192, 768), 'float16', 'float16', layout1='NHTD', layout2='NHTD', layout_out='NHDT',

  100.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  101.                 dim="0 0 64 64 0 1 64 64 0 2 64 4", bind_block="12 12", bind_thread="32 4")

  102.     test_ms_bmm((64, 12, 128, 128), (64, 12, 128, 64), 'float16', 'float16', layout1='NHDT', layout2='NHTD', layout_out='NHDT',

  103.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  104.                 dim="0 0 1 1 0 1 1 1 0 2 64 64 0 3 64 64 0 4 64 4", bind_block="12 64", bind_thread="32 4")

  105.     test_ms_bmm((64, 12, 128, 64), (64, 12, 128, 64), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  106.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  107.                 dim="0 0 1 1 0 1 1 1 0 2 64 64 0 1 128 128 0 2 64 4", bind_block="2 12 64", bind_thread="32 4")

  108.     test_ms_bmm((8192, 768), (8192, 3072), 'float16', 'float16', layout1='NHTD', layout2='NHTD', layout_out='NHDT',

  109.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  110.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="24 6", bind_thread="32 8")

  111.     test_ms_bmm((8192, 3072), (8192, 768), 'float16', 'float16', layout1='NHTD', layout2='NHTD', layout_out='NHDT',

  112.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  113.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 24", bind_thread="32 8")

  114.     test_ms_bmm((8192, 768), (3072, 768), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  115.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  116.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="24 64", bind_thread="32 4")

  117.     test_ms_bmm((8192, 3072), (768, 3072), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  118.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  119.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 64", bind_thread="32 8")

  120.     test_ms_bmm((8192, 3072), (3072, 768), 'float16', 'float16', layout1='NHDT', layout2='NHTD', layout_out='NHDT',

  121.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  122.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 64", bind_thread="32 8")

  123.     test_ms_bmm((8192, 768), (768, 3072), 'float16', 'float16', layout1='NHDT', layout2='NHTD', layout_out='NHDT',

  124.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  125.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="24 64", bind_thread="32 4")

  126.     """

  127.     

  128.     """ Bert Batch32

  129.     test_ms_bmm((4096, 768), (768, 768), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  130.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  131.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 32", bind_thread="32 4")

  132.     test_ms_bmm((32, 12, 128, 128), (32, 12, 128, 64), 'float16', 'float16', layout1='NHDT', layout2='NHTD', layout_out='NHDT',

  133.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  134.                 dim="0 0 1 1 0 1 1 1 0 2 64 64 0 3 64 64 0 4 64 4", bind_block="12 32", bind_thread="32 4")

  135.     test_ms_bmm((4096, 768), (4096, 768), 'float16', 'float16', layout1='NHTD', layout2='NHTD', layout_out='NHDT',

  136.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  137.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 6", bind_thread="32 4")

  138.     test_ms_bmm((32, 12, 128, 64), (32, 12, 128, 64), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  139.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  140.                 dim="0 0 1 1 0 1 1 1 0 2 32 32 0 3 64 64 0 4 32 4", bind_block="12 32", bind_thread="32 4")

  141.     test_ms_bmm((4096, 768), (4096, 3072), 'float16', 'float16', layout1='NHTD', layout2='NHTD', layout_out='NHDT',

  142.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  143.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="24 6", bind_thread="32 8")

  144.     test_ms_bmm((4096, 3072), (768, 3072), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  145.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  146.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 32", bind_thread="32 8")

  147.     test_ms_bmm((4096, 3072), (4096, 768), 'float16', 'float16', layout1='NHTD', layout2='NHTD', layout_out='NHDT',

  148.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  149.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 24", bind_thread="32 8")

  150.     test_ms_bmm((4096, 3072), (3072, 768), 'float16', 'float16', layout1='NHDT', layout2='NHTD', layout_out='NHDT',

  151.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  152.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="6 32", bind_thread="32 4")

  153.     test_ms_bmm((4096, 768), (3072, 768), 'float16', 'float16', layout1='NHDT', layout2='NHDT', layout_out='NHDT',

  154.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  155.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="24 32", bind_thread="32 4")

  156.     test_ms_bmm((4096, 768), (768, 3072), 'float16', 'float16', layout1='NHDT', layout2='NHTD', layout_out='NHDT',

  157.                 shape_bias=(1, ), add_bias=False, tensor_core=True, poly_sch=poly_sch, 

  158.                 dim="0 0 128 128 0 1 128 128 0 2 64 4", bind_block="24 32", bind_thread="32 4")

  159.      """

  160. 

  161. def cast(poly_sch, fuzz_shape=None, mind_trick_str=''):

  162.     test_ms_cast((32, 32, 14, 14, 16), "float16", "float32", poly_sch=poly_sch)

  163.     test_ms_cast((32, 32, 14, 14, 16), "float32", "float16", poly_sch=poly_sch)

  164. 

  165. 

  166. def exp(poly_sch, fuzz_shape=None, mind_trick_str=''):

  167.     test_ms_exp((1024, 4096), 'float32', poly_sch=poly_sch)

  168.     test_ms_exp((1024, 4096), 'float16', poly_sch=poly_sch)

  169.     test_ms_exp((1024, 4095), 'float16', poly_sch=poly_sch)

  170.     test_ms_exp((1024, 799), 'float16', poly_sch=poly_sch)

  171. 

  172. 

  173. def maximum(poly_sch, fuzz_shape=None, mind_trick_str=''):

  174.     test_ms_maximum((32, 1024, 1024), (32, 1024, 1024),

  175.                     'float32', poly_sch=poly_sch)

  176.     test_ms_maximum((32, 1024, 1024), (1, 1024, 1024),

  177.                     'float16', poly_sch=poly_sch)

  178.     test_ms_maximum((32, 32, 32, 256), (32, 32, 1, 256),

  179.                     'float16', poly_sch=poly_sch)

  180. 

  181. 

  182. def minimum(poly_sch, fuzz_shape=None, mind_trick_str=''):

  183.     test_ms_minimum((32, 1024, 1024), (32, 1024, 1024),

  184.                     'float32', poly_sch=poly_sch)

  185.     test_ms_minimum((32, 1024, 1024), (1, 1024, 1024),

  186.                     'float16', poly_sch=poly_sch)

  187.     test_ms_minimum((32, 32, 32, 256), (32, 32, 1, 256),

  188.                     'float16', poly_sch=poly_sch)

  189. 

  190. 

  191. def mul(poly_sch, fuzz_shape=None, mind_trick_str=''):

  192.     test_ms_mul((1024, 4096), 'float32', poly_sch=poly_sch)

  193. 

  194. 

  195. def divide(poly_sch, fuzz_shape=None, mind_trick_str=''):

  196.     test_ms_divide((1024, 1024), 'float32', poly_sch=poly_sch)

  197.     test_ms_divide((1024, 1024), 'float16', poly_sch=poly_sch)

  198. 

  199. 

  200. def reshape(poly_sch, fuzz_shape=None, mind_trick_str=''):

  201.     test_ms_reshape("float32", (64, 128, 1024),

  202.                     (8192, 1024), poly_sch=poly_sch)

  203.     test_ms_reshape("float16", (64, 128, 1024),

  204.                     (8192, 1024), poly_sch=poly_sch)

  205. 

  206. 

  207. def rsqrt(poly_sch, fuzz_shape=None, mind_trick_str=''):

  208.     test_ms_rsqrt((32, 1024, 1024), 'float32', poly_sch=poly_sch)

  209.     test_ms_rsqrt((32, 1024, 1024), 'float16', poly_sch=poly_sch)

  210. 

  211. 

  212. def sqrt(poly_sch, fuzz_shape=None, mind_trick_str=''):

  213.     test_ms_sqrt((1024, 1024), "float32", poly_sch=poly_sch)

  214.     test_ms_sqrt((1024, 1024), "float16", poly_sch=poly_sch)

  215. 

  216. 

  217. def sub(poly_sch, fuzz_shape=None, mind_trick_str=''):

  218.     test_ms_sub((32, 1024, 1024), (32, 1024, 1024),

  219.                 'float32', poly_sch=poly_sch)

  220.     test_ms_sub((32, 1024, 1024), (32, 1024, 1024),

  221.                 'float16', poly_sch=poly_sch)

  222.     test_ms_sub((32, 1024, 1024), (1, 1024, 1024),

  223.                 'float32', poly_sch=poly_sch)

  224.     test_ms_sub((4, 4, 4), (1, 4, 4), 'float32', poly_sch=poly_sch)

  225. 

  226. 

  227. def tile(poly_sch, fuzz_shape=None, mind_trick_str=''):

  228.     test_ms_tile((1024, 4096), (3,), 'float32', poly_sch=poly_sch)

  229.     test_ms_tile((1024, 4096), (3,), 'float16', poly_sch=poly_sch)

  230. 

  231. 

  232. def one_hot(poly_sch, fuzz_shape=None, mind_trick_str=''):

  233.     test_ms_one_hot((1024,), 16, "int32", 1, 0, 0, poly_sch=poly_sch)

  234.     test_ms_one_hot((1024,), 16, "float32", 1, 0, 0, poly_sch=poly_sch)

  235.     test_ms_one_hot((32,), 16, "int32", 1, 0, 0, poly_sch=poly_sch)

  236.     test_ms_one_hot((32,), 16, "float32", 1, 0, 0, poly_sch=poly_sch)

  237. 

  238. 

  239. def expand_dims(poly_sch, fuzz_shape=None, mind_trick_str=''):

  240.     test_expand_dims((32, 1024, 1024), 1, 'float32', poly_sch=poly_sch)

  241.     test_expand_dims((32, 1024, 1024), 2, 'float16', poly_sch=poly_sch)

  242. 

  243. 

  244. def trans_data(poly_sch, fuzz_shape=None, mind_trick_str=''):

  245.     test_ms_trans_data((8, 24, 38, 38), (0, 2, 1, 3),

  246.                        'float32', poly_sch=poly_sch)

  247.     test_ms_trans_data((8, 24, 38, 38), (0, 2, 1, 3),

  248.                        'float16', poly_sch=poly_sch)

  249. 

  250. 

  251. def log(poly_sch, fuzz_shape=None, mind_trick_str=''):

  252.     test_ms_log((9, 1024, 1024), 'float16', poly_sch=poly_sch)

  253.     test_ms_log((9, 1024, 1024), 'float32', poly_sch=poly_sch)

  254. 

  255. 

  256. def pow(poly_sch, fuzz_shape=None, mind_trick_str=''):

  257.     test_ms_pow((9, 1024, 1024), (9, 1024, 1024), 'float32', poly_sch=poly_sch)

  258.     test_ms_pow((9, 1024, 1024), (9, 1024, 1), 'float32', poly_sch=poly_sch)

  259.     test_ms_pow((9, 1024, 1024), (9, 1, 1), 'float32', poly_sch=poly_sch)

  260.     test_ms_pow((9, 1024, 1024), (1, 1, 1), 'float32', poly_sch=poly_sch)

  261.     test_ms_pow((9, 1024, 1024), (9, 1024, 1024), 'float16', poly_sch=poly_sch)

  262.     test_ms_pow((9, 1024, 1024), (9, 1024, 1), 'float16', poly_sch=poly_sch)

  263.     test_ms_pow((9, 1024, 1024), (9, 1, 1), 'float16', poly_sch=poly_sch)

  264.     test_ms_pow((9, 1024, 1024), (1, 1, 1), 'float16', poly_sch=poly_sch)

  265. 

  266. 

  267. def abs(poly_sch, fuzz_shape=None, mind_trick_str=''):

  268.     test_ms_abs((1024, 1024), "float32", poly_sch=poly_sch)

  269.     test_ms_abs((1024, 1024), "float16", poly_sch=poly_sch)

  270.     test_ms_abs((1, ), "float32", poly_sch=poly_sch)

  271.     test_ms_abs((1, 1), "float32", poly_sch=poly_sch)

  272.     test_ms_abs((1, ), "float16", poly_sch=poly_sch)

  273.     test_ms_abs((1, 1), "float16", poly_sch=poly_sch)

  274. 

  275. 

  276. def neg(poly_sch, fuzz_shape=None, mind_trick_str=''):

  277.     test_ms_neg((1024, 1024), "float32", poly_sch=poly_sch)

  278.     test_ms_neg((1024, 1024), "float16", poly_sch=poly_sch)

  279.     test_ms_neg((1, ), "float32", poly_sch=poly_sch)

  280.     test_ms_neg((1, 1), "float32", poly_sch=poly_sch)

  281.     test_ms_neg((1, ), "float16", poly_sch=poly_sch)

  282.     test_ms_neg((1, 1), "float16", poly_sch=poly_sch)

  283. 

  284. 

  285. def round(poly_sch, fuzz_shape=None, mind_trick_str=''):

  286.     test_ms_round((1024, 1024), "float32", poly_sch=poly_sch)

  287.     test_ms_round((1024, 1024), "float16", poly_sch=poly_sch)

  288.     test_ms_round((1, ), "float32", poly_sch=poly_sch)

  289.     test_ms_round((1, 1), "float32", poly_sch=poly_sch)

  290.     test_ms_round((1, ), "float16", poly_sch=poly_sch)

  291.     test_ms_round((1, 1), "float16", poly_sch=poly_sch)

  292. 

  293. 

  294. def reduce_sum(poly_sch, fuzz_shape=None, mind_trick_str=''):

  295.     test_ms_reduce_sum((256, 256), 'float32', axis=(1,),

  296.                        keepdims=True, poly_sch=poly_sch)

  297.     test_ms_reduce_sum((9, 1024, 1024), 'float32', axis=None,

  298.                        keepdims=False, poly_sch=poly_sch)

  299.     test_ms_reduce_sum((9, 1024, 1024), 'float32', axis=2,

  300.                        keepdims=True, poly_sch=poly_sch)

  301.     test_ms_reduce_sum((9, 1024), 'float16', axis=None,

  302.                        keepdims=False, poly_sch=poly_sch)

  303.     test_ms_reduce_sum((9, 1024), 'float16', axis=1,

  304.                        keepdims=True, poly_sch=poly_sch)

  305. 

  306. 

  307. def select(poly_sch, fuzz_shape=None, mind_trick_str=''):

  308.     test_ms_select((2, ), (2, 2, 2),  "int8", "float16", poly_sch=poly_sch)

  309. 

  310. 

  311. def equal(poly_sch, fuzz_shape=None, mind_trick_str=''):

  312.     test_ms_equal(((1, 1024), (1, 1024)), 'float16', poly_sch=poly_sch)

  313.     test_ms_equal(((1, 1024), (1, 1024)), 'float32', poly_sch=poly_sch)

  314. 

  315. 

  316. def less_equal(poly_sch, fuzz_shape=None, mind_trick_str=''):

  317.     test_ms_less_equal((1, 1024), (1, 1024), 'float16', poly_sch=poly_sch)

  318.     test_ms_less_equal((1, 1024), (1, 1024), 'float32', poly_sch=poly_sch)

  319. 

  320. 

  321. def greater_equal(poly_sch, fuzz_shape=None, mind_trick_str=''):

  322.     test_ms_greater_equal((1, 1024), (1, 1024), 'float16', poly_sch=poly_sch)

  323.     test_ms_greater_equal((1, 1024), (1, 1024), 'float32', poly_sch=poly_sch)

  324. 

  325. 

  326. def reciprocal(poly_sch, fuzz_shape=None, mind_trick_str=''):

  327.     test_ms_reciprocal((1, 1024), 'float16', poly_sch=poly_sch)

  328.     test_ms_reciprocal((1, 1024), 'float32', poly_sch=poly_sch)

  329. 

  330. 

  331. def reduce_min(poly_sch, fuzz_shape=None, mind_trick_str=''):

  332.     test_ms_reduce_min((9, 1024, 1024), 'float32', axis=None,

  333.                        keepdims=False, poly_sch=poly_sch)

  334.     test_ms_reduce_min((9, 1024, 1024), 'float16', axis=None,

  335.                        keepdims=False, poly_sch=poly_sch)

  336.     test_ms_reduce_min((9, 1024, 1024), 'float32', axis=2,

  337.                        keepdims=False, poly_sch=poly_sch)

  338.     test_ms_reduce_min((9, 1024, 1024), 'float16', axis=2,

  339.                        keepdims=False, poly_sch=poly_sch)

  340. 

  341. 

  342. def reduce_max(poly_sch, fuzz_shape=None, mind_trick_str=''):

  343.     test_ms_reduce_max((9, 1024, 1024), 'float32', axis=None,

  344.                        keepdims=False, poly_sch=poly_sch)

  345.     test_ms_reduce_max((9, 1024, 1024), 'float16', axis=None,

  346.                        keepdims=False, poly_sch=poly_sch)

  347.     test_ms_reduce_max((9, 1024, 1024), 'float32', axis=2,

  348.                        keepdims=False, poly_sch=poly_sch)

  349.     test_ms_reduce_max((9, 1024, 1024), 'float16', axis=2,

  350.                        keepdims=False, poly_sch=poly_sch)

  351. 

  352. 

  353. def fused_pad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  354.     test_fused_pad((7, 7, 3, 64), (0, 0, 0, 0), (0, 0, 1, 0),

  355.                    layout='NHWC', pad_value=0.0, poly_sch=poly_sch)

  356. 

  357. 

  358. def fused_bn_reduce(poly_sch, fuzz_shape=None, mind_trick_str=''):

  359.     test_fused_bn_reduce((256, 7, 7, 2048), layout='NHWC', poly_sch=poly_sch)

  360. 

  361. 

  362. def fused_bn_update(poly_sch, fuzz_shape=None, mind_trick_str=''):

  363.     test_fused_bn_update((2048,), poly_sch=poly_sch, mind_trick=mind_trick_str)

  364. 

  365. 

  366. def fused_bn_follow_relu(poly_sch, fuzz_shape=None, mind_trick_str=''):

  367.     test_fused_bn_follow_relu(

  368.         (256, 7, 7, 2048), layout='NHWC', poly_sch=poly_sch)

  369. 

  370. 

  371. def fused_bn_follow_relu_avgpool(poly_sch, fuzz_shape=None, mind_trick_str=''):

  372.     test_fused_bn_follow_relu_avgpool(

  373.         (256, 7, 7, 2048), layout='NHWC', poly_sch=poly_sch)

  374. 

  375. 

  376. def fused_bn_double_follow_relu(poly_sch, fuzz_shape=None, mind_trick_str=''):

  377.     test_fused_bn_double_follow_relu(

  378.         (256, 7, 7, 2048), layout='NHWC', poly_sch=poly_sch)

  379. 

  380. 

  381. def fused_bn_reduce_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  382.     test_fused_bn_reduce_grad(

  383.         (256, 56, 56, 256), layout='NHWC', poly_sch=poly_sch)

  384. 

  385. 

  386. def fused_relu_grad_bn_reduce_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  387.     test_fused_relu_grad_bn_reduce_grad(

  388.         (64, ), (256, 112, 112, 64),  layout='NHWC', poly_sch=poly_sch)

  389. 

  390. 

  391. def fused_relu_grad_bn_double_reduce_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  392.     test_fused_relu_grad_bn_double_reduce_grad(

  393.         (256,), (256, 56, 56, 256), layout="NHWC", poly_sch=poly_sch)

  394. 

  395. 

  396. def fused_l2loss_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  397.     test_fused_l2loss_grad((1, 1, 256, 1024), layout='NHWC', poly_sch=poly_sch, mind_trick=mind_trick_str)

  398. 

  399. 

  400. def fused_is_finite(poly_sch, fuzz_shape=None, mind_trick_str=''):

  401.     test_fused_is_finite((1, 1, 256, 1024), layout='NHWC', poly_sch=poly_sch)

  402. 

  403. 

  404. def fused_relu_grad_bn_update_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  405.     test_fused_relu_grad_bn_update_grad(

  406.         (256, 112, 112, 64), (64,), layout="NHWC", poly_sch=poly_sch)

  407. 

  408. 

  409. def fused_relu_grad_bn_double_update_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  410.     test_fused_relu_grad_bn_double_update_grad(

  411.         (256, 56, 56, 256), (256, ), layout='NHWC', poly_sch=poly_sch)

  412. 

  413. 

  414. def fused_relu_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  415.     test_fused_relu_grad((256, 56, 56, 256), poly_sch=poly_sch, mind_trick=mind_trick_str)

  416. 

  417. 

  418. def fused_bn_update_grad(poly_sch, fuzz_shape=None, mind_trick_str=''):

  419.     test_fused_bn_update_grad(

  420.         (256, 56, 56, 256), (256,), layout="NHWC", poly_sch=poly_sch)

  421. 

  422. 

  423. def fused_mul_div_rsqrt_mul_isfinite_red(poly_sch, fuzz_shape=None, mind_trick_str=''):

  424.     test_fused_mul_div_rsqrt_mul_isfinite_red((64,), poly_sch=poly_sch)

  425. 

  426. 

  427. class Logger(object):

  428.     def __init__(self, filename, stream):

  429.         self.terminal = stream

  430.         self.log = open(filename, 'a')

  431. 

  432.     def write(self, message):

  433.         self.terminal.write(message)

  434.         self.log.write(message)

  435. 

  436.     def flush(self):

  437.         pass

  438. 

  439. 

  440. def usage(op_map):

  441.     print("Usage:")

  442.     print("1. Run func1 and func2 with manual schedule:")

  443.     print("\t$python test_all.py -m func_name1 func_name2")

  444.     print("\t$python test_all.py --manual func_name1 func_name2")

  445.     print("2. Run all with auto schedule:")

  446.     print("\t$python test_all.py -a all\n")

  447.     print("\t$python test_all.py --auto all\n")

  448.     print("3. Both schedule methods will be tested if no option is specified")

  449.     print("\t$python test_all.py func_name")

  450.     print("4. Run fuzz test of add op with maximal dimension of shape equals 3")

  451.     print("\t$python test_all.py -f 3 add")

  452.     print("Available func:")

  453.     print("\t", list(op_map.keys()), "\n")

  454. 

  455. 

  456. if __name__ == '__main__':

  457.     import sys

  458.     import getopt

  459.     import traceback

  460.     from datetime import datetime

  461. 

  462.     op_map = {"abs": abs, "add": add, "addn": addn, "bmm": bmm, "cast": cast, "divide": divide,

  463.               "equal": equal, "exp": exp, "greater_equal": greater_equal, "less_equal": less_equal,

  464.               "log": log, "max": maximum, "min": minimum, "mul": mul, "neg": neg, "pow": pow,

  465.               "reciprocal": reciprocal, "round": round, "rsqrt": rsqrt, "select": select, "sqrt": sqrt,

  466.               "sub": sub, "reduce_max": reduce_max, "reduce_min": reduce_min,

  467.               "reduce_sum": reduce_sum, "expand_dims": expand_dims, "one_hot": one_hot,

  468.               "reshape": reshape, "tile": tile, "trans_data": trans_data,

  469.               "fused_pad": fused_pad,

  470.               "fused_bn_reduce": fused_bn_reduce,

  471.               "fused_bn_update": fused_bn_update,

  472.               "fused_bn_follow_relu": fused_bn_follow_relu,

  473.               "fused_bn_follow_relu_avgpool": fused_bn_follow_relu_avgpool,

  474.               "fused_bn_double_follow_relu": fused_bn_double_follow_relu,

  475.               "fused_bn_reduce_grad": fused_bn_reduce_grad,

  476.               "fused_relu_grad_bn_reduce_grad": fused_relu_grad_bn_reduce_grad,

  477.               "fused_relu_grad_bn_double_reduce_grad": fused_relu_grad_bn_double_reduce_grad,

  478.               "fused_l2loss_grad": fused_l2loss_grad,

  479.               "fused_is_finite": fused_is_finite,

  480.               "fused_relu_grad_bn_update_grad": fused_relu_grad_bn_update_grad,

  481.               "fused_relu_grad_bn_double_update_grad": fused_relu_grad_bn_double_update_grad,

  482.               "fused_relu_grad": fused_relu_grad,

  483.               "fused_bn_update_grad": fused_bn_update_grad,

  484.               "fused_mul_div_rsqrt_mul_isfinite_red": fused_mul_div_rsqrt_mul_isfinite_red

  485.               }

  486.     all_f = list(op_map.values())

  487.     op_map["all"] = all_f

  488.     if len(sys.argv) == 1:

  489.         usage(op_map)

  490.         sys.exit()

  491. 

  492.     options, args = getopt.getopt(

  493.         sys.argv[1:], "f:", ["fuzz=", "mind-trick-string=", "mind-trick-file="])

  494.     mind_trick_str = ''

  495.     fuzz_dim = 0

  496.     for name, value in options:

  497.         if name in ("-f", "--fuzz"):

  498.             fuzz_dim = int(value)

  499.         if name == "--mind-trick-string":

  500.             mind_trick_str = value

  501.         if name == "--mind-trick-file":

  502.             with open(value, 'r') as f:

  503.                 mind_trick_str = f.read()

  504. 

  505.     fail_op_list = dict()

  506.     run_op_list = list()

  507.     for op in args:

  508.         if op_map.get(op) is not None:

  509.             f = op_map.get(op)

  510.             if not isinstance(f, list):

  511.                 run_op_list.append(f)

  512.             else:

  513.                 run_op_list += f

  514. 

  515.     now = datetime.now()

  516.     filename = "opstest_" + \

  517.         '-'.join(list(map(str, [now.month, now.day,

  518.                                 now.hour, now.minute]))) + ".log"

  519.     sys.stdout = Logger(filename, sys.stdout)

  520.     sys.stderr = Logger(filename, sys.stderr)

  521. 

  522.     for op in run_op_list:

  523.         print("Operater: ", op.__name__)

  524.         fuzz_shape = gen_random_shape(fuzz_dim) if fuzz_dim > 0 else None

  525.         if fuzz_shape:

  526.             print("Fuzz shape: {}".format(fuzz_shape))

  527.         try:

  528.             print("Time of auto schedule:")

  529.             op(poly_sch=True, fuzz_shape=fuzz_shape, mind_trick_str=mind_trick_str)

  530.         except:

  531.             if op.__name__ in fail_op_list:

  532.                 fail_op_list[op.__name__].extend(

  533.                     ["using auto schedule:", traceback.format_exc()])

  534.             else:

  535.                 fail_op_list[op.__name__] = [

  536.                     "using auto schedule:", traceback.format_exc()]

  537. 

  538.     if len(fail_op_list) == 0:

  539.         print("All test pass!")

  540.     else:

  541.         for op, error_info in fail_op_list.items():

  542.             print("Run op %s error" % op)

  543.             for e in error_info:

  544.                 print(e)